1

Iconicity in Expressives: An Empirical

Investigation
MICHAEL GASSER, NITYA SETHURAMAN, AND STEPHEN
HOCKEMA

1 Introduction
All comprehensive linguistic theories must confront the nature of the rela-
tionship between linguistic forms and the functions they perform. To what
extent is this relationship constrained and where do the constraints come
from? At the level of grammar, form-function relationships are apparently
constrained in part by compositionality, and a range of linguistic or cogni-
tive universals governing grammatical mappings have also been proposed.
At the level of the lexicon, associations between form and function have, at
least since Saussure (1966), been assumed to be mostly unconstrained. That
is, the word form expressing a particular meaning in a particular language
could be any phonotactically legal sequence of phonological segments and
the convention for each such relationship must simply be memorized by
learners of the language. It is in this sense that the form-function relation-
ship in language is said to be mainly “arbitrary”.
At the same time, it has often been recognized that within restricted re-
gions of the lexicon, complete arbitrariness fails to hold. This is most obvi-
ous for onomatopoeia (e.g., pitter patter, moo), but it has also been attested
for other categories of words in a wide range of spoken and signed lan-
guages (Hinton, Nichols & Ohala, 1994; Taub, 2001). In all of these cases,
the forms of words are said to somehow “suggest” their meanings, or the

Experimental and Empirical Methods.

Sally Rice and John Newman (eds.).
Copyright © 2005, CSLI Publications.
1
2 / MICHAEL GASSER, NITYA SETHURAMAN, AND STEPHEN HOCKEMA

meanings are said to somehow “motivate” the word forms. The general
term used for these phenomena is iconicity. Despite increased interest in
iconicity in recent years, two fundamental questions remain unanswered:
1. What empirical means do we have for identifying a particular word
as iconic or arbitrary?
2. How is it that the lexicon can be both arbitrary and iconic?
A possible answer to the second question is suggested by a completely
different body of research, the study of lexical access and word recognition
within psycholinguistics. Here it is recognized that the similarity of a word
form to other forms plays a role in the speed at which the word is accessed
during comprehension or production (Luce & Pisoni, 1998). Similarly, re-
search on semantic speech errors during production (e.g., Fromkin, 1971)
shows that words with similar meanings may interfere with one another.
Since phonological similarity and semantic similarity both have an effect on
language processing, it is worth considering whether the relationship be-
tween the two types of similarity also plays a role and whether communica-
tive considerations could govern the sorts of form-meaning relationships
that are favored or disfavored in languages.
In this paper, we provide a preliminary method of measuring the iconic-
ity of words. This method relies on a new, formal definition of iconicity,
one that distinguishes between what we call absolute iconicity, based on the
direct relationship between form and meaning, and relative iconicity, based
on the relationship between distance between forms and distance between
meanings. We also propose an account of how arbitrariness and iconicity
might be favored for different parts of the lexicon. We describe a prelimi-
nary study using Japanese and Tamil expressives, a lexical category report-
edly characterized by a high degree of iconicity (Childs, 1994).

2 Expressives
Many languages of the world, especially in Africa, South Asia, Southeast
Asia, and the Americas, have a lexical category referred to as expressives,
also sometimes as ideophones or mimetics (Childs, 1994). While there are
no necessary and sufficient conditions for defining this category cross-
linguistically, a number of phonological, morphological, syntactic, semantic,
and pragmatic properties characterize the expressive prototype. Expressive
forms tend to be made up of a limited set of phonemes and to exhibit redu-
plication. Syntactically, expressives are often used in combination with a
‘light’ verb such as say or do. Expressives tend to denote adverbial mean-
ings, and among the semantic categories often expressed by this category
are movements (of the body or of objects); physical states; sounds and
noises; speech patterns; sensations, emotions or mental states; personal ap-
 3 / ICONICITY IN EXPRESSIVES: AN EMPIRICAL INVESTIGATION

pearance; facial expressions; and personality traits. In this paper, we focus

on expressives in two unrelated languages, Japanese and Tamil, specifically
on the extent to which they exhibit iconicity. Some examples of expressives
in these languages are given in Table 1.
Table 1: Examples of Japanese and Tamil expressives

Japanese Tamil
kirakira ‘glitter, twinkle’ labakk ‘suddenly’
giragira ‘glare strongly’ pacakk ‘suddenly’
girari ‘glare momentarily’ didikku ‘all of a sudden’

If it is true, as often claimed (e.g., Childs, 1994), that expressives are

characterized by iconicity, why would this be? And do languages with ex-
pressives have other categories of words that are clearly distinguishable
from expressives in terms of their arbitrariness? We return to these ques-
tions below. But first we need to be more precise about what we mean by
“iconicity”. What is it about words such as those in Table 1 that makes them
iconic, and how would we demonstrate their iconicity in an objective man-
ner?

3 Types of Iconicity
We consider words to be associations of what we will call forms and mean-
ings. Both forms and meanings can be thought of as points in multi-
dimensional spaces, though we may not have direct access to what the set of
defining dimensions is. Given a meaning, a language user should be able to
assign a form to it: this is production. Given a form, a language user should
be able to assign a meaning to it: this is comprehension. Figure 1 illustrates
these two processes for the simple case of two form and two meaning di-
mensions.

Figure 1. Production and Comprehension. Form space (F) and Meaning space (M) are each
defined over only two dimensions. Each small square represents the form or meaning of a word.
In production, the speaker begins with a meaning (m1) and accesses a form (f1). In comprehen-
sion, the addressee begins with a form (f1) and accesses a meaning (m1).
 We define absolute iconicity to be the property of a set of words for
which there is a similarity function that relates form and meaning (either in
whole or part). In absolute iconicity, there is a correlation between one or
more form dimensions and one or more meaning dimensions. The most
obvious example is onomatopoeia, for which forms are intended as imita-
tions of sounds in nature. Thus there appears to be a weak correlation be-
tween the vowel formants in conventional words for animal sounds (moo,
quack, cheep) and the perceived formants in the sounds made by the ani-
mals. Sign languages offer many other instances of absolute iconicity, for
example, the conventional use of the space in front of the signer for future
times and the space behind the signer for past times. Note that the form-
meaning function itself can vary from one language to another. Though we
know of no such languages, it is at least conceivable that a sign language
could assign the future to the space behind the signer and the past to the
space in front of the signer. The point is that the assignment applies consis-
tently across all of the words in the iconic set.
We define relative iconicity to be the property of a set of words for
which there is a correlation between form similarity and meaning similarity.
That is, rather than a similarity function relating meaning to form, relative
iconicity is based on separate similarity functions relating forms to forms
and meanings to meanings. For example, glow, gleam, glint, and glimmer
seem to exhibit relative iconicity because their similarity of form (gl- in the
salient initial position) corresponds to a similarity of meaning (‘having to do
with light of low intensity’) and because there are relatively few other Eng-
lish verbs that begin with gl- and do not share this meaning. Note that with
relative iconicity there need not be form-meaning similarity – this relation-
ship could be completely arbitrary – as long as similar forms have similar
meanings and similar meanings have similar forms. Note also that both
kinds of iconicity are properties of groups of words, absolute iconicity be-
cause it is defined with respect to a similarity function over a set of form-
meaning pairings and relative iconicity because it is defined with respect to
the similarity between pairs within a group of words. In particular it makes
no sense to say that a particular word is relatively iconic.
We define arbitrariness to be the absence of either absolute or relative
iconicity. In this paper we will be mainly concerned with arbitrariness in the
latter sense. For example, the words in the set apple, pear, peach, apricot,
plum, and cherry appear to exhibit arbitrariness in the sense of absence of
relative iconicity within the set since there is no obvious relationship be-
tween the similarity of forms and the prima facie similarity of associated
meanings. For example, peaches and apricots are quite similar (relative to
other objects or foods), but the forms of the words are as different as any
 5 / ICONICITY IN EXPRESSIVES: AN EMPIRICAL INVESTIGATION

two pairs on this list. And the meanings of the only two nouns in the list
that begin with vowels (in both cases followed by /p/) appear no more simi-
lar than any other pair.
A special kind of relative iconicity is what we call anti-iconicity, where
expressions with similar meanings tend to have dissimilar forms. That is,
there is a negative correlation between form similarity and meaning similar-
ity. One of the goals of this paper is to demonstrate that anti-iconicity actu-
ally exists within some limited domains, but we illustrate here with a trivial
example. Consider these common nouns for carnivores: dog, wolf, fox, cat.
If we were to judge the similarities of the meanings of these words, relying
either on their zoological classification or on their gross appearance and
behavioral patterns1, we might find dogs and wolves most similar and foxes
more similar to dogs/wolves than they are to cats. But with respect to pho-
netic form, we find that cat and dog both consist of stop-vowel-stop se-
quences, whereas both wolf and fox consist of consonant-vowel-consonant-
fricative sequences. That is, it would appear that within this set of words,
form similarity correlates negatively with meaning similarity.
One problem with our informal examples of relative iconicity, arbitrari-
ness, and anti-iconicity is that they have relied on our intuitive impressions
of form and meaning similarity. For the definitions to be useful, we need
explicit means of measuring form similarity and meaning similarity. With
no constraints on what counts as similar, one could argue, for example, that
in our fruit example, the presence of /p/ in words for fruits in the list that do
not grow in clusters is evidence for iconicity within this list. We return to
this issue in Section 6 below, offering a formal measure of form similarity
and tentative approach to the trickier problem of meaning similarity.
In Figure 2, we abstractly illustrate absolute and relative iconicity and in
Figure 3, anti-iconicity and arbitrariness.

1 Note that these are not the only possible criteria for comparing the animals. On a scale of
tameness, for example, the similarities would clearly turn out differently.
Figure 2. Absolute and Relative Iconicity. The arrows indicate associations between the forms
and meanings of individual words. For the absolute iconicity example, there is a perfect corre-
lation between both pairs of F and M dimensions. For the relative iconicity example, we high-
light the connections for one cluster of words to show how similarity of form is associated with
similarity of meaning. In the plot on the right, each point represents the form and meaning
similarities for a single pair of words in the iconic domain.

Figure 3. Relative Anti-Iconicity and Arbitrariness. For the anti-iconic example, one cluster of
meanings is highlighted to show how its forms tend to be further apart than they would be by
chance. As in Figure 2, in the plots on the right, each point represents the form and meaning
similarities for a pair of words in the domain.
 7 / ICONICITY IN EXPRESSIVES: AN EMPIRICAL INVESTIGATION

Given these definitions of relative iconicity and anti-iconicity, there are

three possibilities for what is actually true of a particular natural language
(different languages could differ in this regard):
1. The lexicon of the language could be arbitrary. Any apparent rela-
tive iconicity within a group of words would be spurious; if we
randomly selected enough groups, we would eventually find what
appeared to be iconicity in at least one of them.
2. The lexicon of the language could be relatively iconic. Any appar-
ent arbitrariness within a group of words would be spurious; if we
randomly selected enough groups, we would eventually find what
appeared to be arbitrariness in at least one of them.
3. The lexicon could be arbitrary in some semantic or phonological
regions and relatively iconic in others. This would require some
definition of a region that is not circular; it would not be enough to
show that a language exhibits iconicity for one set of semantically
related words and arbitrariness for another set because one or the
other could have happened by chance. That is, there would have to
be some independent way of defining the regions where we find
iconicity and those where we find arbitrariness.
We believe that this third possibility is likely, at least for certain languages,
and in the rest of this paper we attempt to show that relative iconicity char-
acterizes expressives and arbitrariness or anti-iconicity characterizes con-
crete nouns. But first we ask why this would be the case.

4 Advantages and Disadvantages of Iconicity and Arbi-

trariness
If we are right and the lexicons of natural languages exhibit both relative
iconicity and arbitrariness, even anti-iconicity, why would this be? Are
there areas within the lexicon that are more compatible with one or the other
sort of relationship and, if so, is there a functional or cognitive explanation
for this difference? In this section, we discuss several possible advantages
and disadvantages of iconicity and arbitrariness for learning, production,
and comprehension.
Iconicity would seem to have an advantage when it comes to the storage
of forms in long-term memory and their retrieval during production. If the
learner already knows one or more words in a domain that is known to be
characterized by relative iconicity, then a new word in that domain should
be easier to remember and to retrieve during production because of the
similarity of its form to forms already learned. That is, if one knows that
form f1 has meaning m1 and is then presented with a new form-meaning pair
f2-m2, it should be easier to remember this pair knowing that the similarity
between f1 and f2 is reflected in the similarity between m1 and m2. It should
also be easier to recall f2 given meaning m2 in production. For a possible
advantage of arbitrariness in learning, see Gasser (2004).
For comprehension, we distinguish two different scenarios in which
relative iconicity would be expected to have different consequences. Con-
sider first a communicative situation in which the addresser (speaker or
signer, S) has produced a word that the addressee (A) is unfamiliar with or
has forgotten. If the word has phonetic properties that mark it as a potential
member of the class of iconic words in the language, for example, if it fea-
tures reduplication, then A may be able to narrow down the range of possi-
ble meanings for the unfamiliar form based on similar, familiar forms. For
example, if A knows that betabeta means ‘sticky’ in Japanese, A may guess
(correctly in this case) that the unfamiliar word betobeto has a similar
meaning. In a context where the range of possible meanings for the unfamil-
iar word is constrained, this help may be all that is needed for A to figure
out what S intended.
Now consider another type of communicative situation, one in which
iconicity could interfere with comprehension. Say the context contains sev-
eral potential referents all belonging to one relatively narrow semantic cate-
gory, for example, gardening tools, fruits, articles of clothing, or internal
organs. S refers to one member of the category with a noun that is familiar
to A; for example, S says, “Go out to the garage and get the hoe.” Now
assume that because of noise of one kind or another, the target word form is
misheard. If the domain were characterized by relative iconicity, then the
misheard form could be understand as a word for another member of the
semantic category, for example, in the sentence above, for a shovel or a rake.
Since the difference between a hoe and a shovel presumably matters in such
a context, relative iconicity has the potential to interfere with communica-
tion. Arbitrariness or anti-iconicity, on the other hand, might have an advan-
tage in such a situation. If the misheard form is a word in the language and
its meaning is very different from that of the intended word, as would be
expected with anti-iconicity, then there is little possibility of confusion with
one of the other potential referents. In the example above, if hoe is misheard
as toe or hole, then A is likely to realize that an error has occurred some-
where since these words don’t refer to objects that one would retrieve from
a garage. Rather than attempt to go and look for a toe or a hole, A would be
expected to ask S for clarification. Compare this with what might happen if
the English word for ‘shovel’ were foe.
Thus we see that both iconicity and arbitrariness or anti-iconicity could
have advantages in communicative situations. Are there areas within the
 9 / ICONICITY IN EXPRESSIVES: AN EMPIRICAL INVESTIGATION

lexicon where one of the other of these situations would be more likely to
arise? We address this in the next section.

5 Where Should Iconicity and Arbitrariness Turn Up?

Specific Predictions
We have argued that arbitrariness or anti-iconicity should be advantageous
when an instance of a word has a set of potential referents that resemble one
another but need to be distinguished. Thus we should expect arbitrariness or
anti-iconicity in semantic domains where there are relatively rigid bounda-
ries between the categories and the distinctions are crucial for communica-
tion. We believe that such semantic domains are most common within the
category of basic concrete nouns: nouns referring to artifacts, to food, to
geographical features.2
On the other hand, we should expect iconicity in semantic domains
where the boundaries between the categories are not crucial for communica-
tion, where the disadvantage of possible confusion between similar mean-
ings does not outweigh the advantage that iconicity should have for permit-
ting the addressee to guess the meanings of unfamiliar words within the
category. We believe that such semantic domains include those that are
typical for expressives in languages that have them, domains such as man-
ner of motion and bodily experience. For example, it may not be particu-
larly important to keep distinct the meanings of groups of words such as
stumble/stagger/hobble and woozy/groggy/dizzy.3
Our predictions then are:
1. Arbitrariness or even anti-iconicity should characterize particular
semantic domains within the category of familiar concrete nouns.
2. Relative iconicity should characterize particular semantic domains
within the category of expressives in languages that have them.
To test these predictions, we require a means of measuring the relative
iconicity for a set of words and a means of isolating a set of concrete nouns
and a set of expressives within candidate semantic categories. In the next
section, we describe how we address each of these requirements.

2 Although we don’t discuss this notion further in this paper, labels also play a role in judg-
ments of similarity (Sloutsky & Fisher, 2004), and anti-iconicity might cause the categories in
such domains to be perceived as more distinct than they would otherwise.
3 We make no claims about whether these English words are iconic by our definition of rela-
tive iconicity. We use them here only to illustrate the sorts of meanings that are conveyed by
expressives in languages such as Japanese and Tamil.
6 Method

6.1 Iconicity Coefficient

We would like to know the relative iconicity for a set of words, that is, the
correlation between the form similarity and meaning similarity for pairs of
words in the category. The measurement of meaning similarity is the more
difficult problem, and for the purposes of this paper, we simplify the prob-
lem by treating words as either inside or outside of particular semantic cate-
gories. This is equivalent to treating meaning similarity as a binary dimen-
sion, with pairs of words within the category in question having maximum
similarity, pairs of words with one member inside and one outside the cate-
gory having minimum similarity, and pairs of words outside the category
having an undefined similarity.
In order to measure the relative iconicity of a category of words, we iso-
late a set of words within the category and a comparable set of words out-
side the category. We then calculate the mean form similarity for pairs of
word forms within the category, sw, and the mean form similarity between
members of the category and words outside the category, sb. Each of these
quantities ranges from close to 0 (completely dissimilar) to 1 (identical).
We define the Iconicity Coefficient (IC) to be sw/sb. If IC is significantly
greater than 1.0 for a category, the category exhibits relative iconicity; if IC
is significantly less than 1.0, the category is anti-iconic; and if IC ≈ 1.0, the
category is arbitrary. The maximum IC value depends on the mean form
similarity between randomly selected pairs of words. Given the algorithm
for form comparison that is outlined below, this quantity is around .05.
That is, as the form similarity of words within the semantic categories ap-
proaches 1, the IC for these very iconic sets approaches 20. Also note that,
since absolute iconicity implies relative iconicity, IC is going to be maximal
for an absolutely iconic language.

6.2 Form Comparison

The calculation of IC requires determining the form similarity of pairs of
words. To accomplish this, we use an algorithm that attempts to align moras
in the two words, yielding a score based on the phonetic similarity of the
phones in corresponding moras and the degree to which the order of moras
in the words is maintained in the alignment.
Each phone is represented as a vector of conventional phonetic features
and phone similarity is 1.0 minus the normalized Euclidian distance be-
tween the phone vectors. The similarity between moras is the normalized
sum of the similarities between corresponding phones in the moras. The
 11 / ICONICITY IN EXPRESSIVES: AN EMPIRICAL INVESTIGATION

relative weight of vowels versus consonants within moras is controlled by a

parameter that we currently set by hand at 0.7. This parameter should
probably depend on the language. The mora alignment algorithm tries all
possible alignments of moras in the two words, counting only the one with
the highest score. The score for an alignment is the mean similarity of the
aligned moras divided by the number of mora pairs in the alignment that
appear in different orders in the two words. A mora in one word with no
corresponding mora in the other word receives a similarity of 0.0. Table 2
shows some example form similarities for pairs of Japanese words and
phone sequences (values range from 0.0 to 1.0).

Table 2: Examples of Japanese form similarity

betabeta / betobeto .762 betabeta / beta .125

betabeta / petapeta .672 betabeta / doron .050
betabeta / gitogito .404 betabeta / pin .016
betabeta / bettari .222

6.3 Selection of Words

For the expressives, we randomly selected 300 words from a dictionary of
expressives for Tamil (Sataasivam, 1966) and an expressive thesaurus for
Japanese (Chang, 1990). For the nouns, we relied on Japanese (Ogura &
Watamaki, 1998) and Tamil (Sethuraman, ms.) versions of the MacArthur-
Bates Communicative Development Inventory (Fenson et al., 1993; hereaf-
ter “CDI”), which lists the words most commonly known by infants and
toddlers. We collected all of the words in the concrete noun categories, to-
taling about 350 items for each language.
For a baseline measure with which to compare results for particular
categories, we calculated mean form similarity within the nouns and expres-
sives and between the two categories for both languages (Table 3). Not sur-
prisingly, in both languages, nouns and expressives are more similar to each
other than to words in the other category. The results also show that Tamil
forms are on average less similar to one another than Japanese forms are,
probably due to the larger phonemic inventory for Tamil.
Table 3: Baseline form similarities

Between nouns
Within nouns Within expressives
and expressives
Japanese .0621 .0920 .0556
Tamil .0427 .0372 .0280
7 Study 1: Iconicity and Anti-Iconicity in Broad Semantic
Categories
We started by isolating relatively broad semantic categories. A thesaurus is
organized in terms of a hierarchical ontology of categories, so, to the extent
that we trust this ontology to accurately reflect how people group concepts,
we can generalize results based on a thesaurus. For this study, we selected a
subset of the most general categories from the Japanese expressive thesau-
rus (Chang, 1990). These included, for example, EMOTION and MOTION.
For the Tamil expressives, we selected roughly comparable categories from
the list of expressives in the Tamil expressive dictionary (Sataasivam, 1966).
For the nouns, we selected several of the high-level categories that group
the words in the CDI, for example, ARTICLES OF CLOTHING and FOOD ITEMS.
Our results show that the Iconicity Coefficient was not significantly
different from 1.0 for any of the categories. That is, within broad categories
grouping expressives and concrete nouns, we found no tendency towards
iconicity or anti-iconicity. Note, however, that our hypotheses for the ad-
vantages of iconicity for expressives and anti-iconicity for concrete nouns
both depend on the degree of semantic similarity within the categories we
are examining. Perhaps the variation within broad categories such as
EMOTION and FOOD ITEMS is too great for these advantages to be reflected.
Given an unfamiliar expressive, there is presumably little advantage for the
hearer to guess that it refers to some emotion on the basis of its similarity to
other emotion words. The possibility of guessing that it refers to more spe-
cific emotions, such as anger or sadness, on the basis of its similarity to
words in those categories, would be more useful. Similarly, there is little
advantage for a misheard word for a food item to be less similar to any
other food word than it is to other nouns. A form for a fruit that could not
easily be confused with the names for other fruits, for example, would be
more useful.
In the next study, we therefore examined narrower semantic categories.

8 Study 2: Iconicity and Anti-Iconicity in Narrow Seman-

tic Categories
In this study we examined the iconicity and anti-iconicity exhibited by the
collected expressives and concrete nouns in meaning categories that are
roughly at a level of specificity corresponding to the narrowest categories in
a thesaurus, for example, words that describe pain of one sort or another. To
collect such categories within the expressives, we selected 19 of the nar-
rowest categories from the Japanese expressive thesaurus (Chang, 1990)
that contained at least 8 words and 11 roughly-corresponding categories
from the Tamil expressive list, using all common words in each category,
 13 / ICONICITY IN EXPRESSIVES: AN EMPIRICAL INVESTIGATION

rather than only those from our original list of 300 expressives. For Japa-
nese these categories included ATTIRE, SLEEP, PAIN, EATING, COMPLAINT,
BURNING, SPLASHING, and STICKINESS; for Tamil they included ANGER,
QUICKNESS, BABBLING , LAUGHTER, EATING, SOFTNESS, and SUDDENNESS.
For the nouns, we informally selected categories roughly comparable in
specificity. We avoided categories in which many of the words were recent
borrowings or were morphologically complex because we felt these proper-
ties could affect the relative iconicity of the words (we return to these ef-
fects in the Conclusion section). Again, we made an attempt to include all
“basic” members of these categories; there were between 8 and 13 of these
in each case. The categories for both languages were FARM ANIMALS,
INTERNAL ORGANS, NATIVE MUSICAL INSTRUMENTS, BODILY EXCRETIONS,
SPICES, and WEATHER PHENOMENA .
To calculate the mean form similarity between words inside and outside
a given category, sb, we used all of the collected expressives and nouns in
each case. For these two languages we take these two categories to be repre-
sentative of the range of forms possible for mono-morphemic words (verbs
in both languages are always inflected). For each category, we wanted to
know whether its Iconicity Coefficient (IC) was significantly greater or less
than 1.0, whether the within-category form similarity differed from the
similarity of those words to the range of all words included in the sample.
We used a Monte Carlo simulation to assess the significance in each case.
This technique works by randomly picking words to place in hypothetical
meaning categories within the larger lexical categories (nouns or expres-
sives) and then using the distribution of the ICs of these categories to de-
termine how likely the ICs we actually obtained would have occurred by
chance.
 Japanese Expressives
TIME **
SECURE ****
STICKY ****
SPLASHING ****
GLITTERING ****
BURNING *
ABRUPT STOP ****
FLOATING ****
CUTTING ****
EATING ***
COMPLAINT ****
ABSENT ***
ANXIETY *
TEARFULNESS ****
PAIN ****
SLEEP
EYE_EXPRESSION **
FAT ****
ATTIRE *

0.6 1.1 1.6 2.1 2.6 3.1 3.6

Iconicity Coefficient

Tamil Expressives
ORNAMENT_NOISE ****

UNEVEN *

SUDDEN ****

SOFT *

HEALTHY
EAT **

HEARTBEAT **

LAUGH
BABBLE ****

QUICK
ANGRY ****

0.6 1.1 1.6 2.1 2.6 3.1 3.6

Iconicity Coefficient

Figure 4. Iconicity Coefficients for Japanese and Tamil Expressives

 15 / ICONICITY IN EXPRESSIVES: AN EMPIRICAL INVESTIGATION

Japanese Nouns

WEATHER

SPICES

EXCRETIONS

MUS.INSTRUMENTS

ORGANS ****

FARM ANIMALS

0.6 1.1 1.6 2.1 2.6 3.1 3.6

Iconicity Coefficient

Tamil Nouns

WEATHER *

SPICES

EXCRETIONS

MUS.INSTRUMENTS *

ORGANS

FARM ANIMALS

0.6 1.1 1.6 2.1 2.6 3.1 3.6

Iconicity Coefficient
Figure 5. Iconicity Coefficients for Japanese and Tamil Nouns and Expressives
The results for expressives and nouns in each language are shown in Fig-
ures 4 and 5. In each figure, the x-axis indicates the IC for each of the cate-
gories of words represented by the horizontal bars. Recall that arbitrariness
corresponds to an IC not significantly different from 1.0, relative iconicity
to an IC significantly greater than 1.0, and anti-iconicity to an IC signifi-
cantly less than 1.0. In the figures, categories with ICs significantly greater
or less than 1.0 are indicated with asterisks next to the bars.4
For the expressives, we see that 8 of the 11 Tamil categories and 18 of
the 19 Japanese categories are iconic. For the nouns, although half of the
categories have ICs of less than one, only 1 of the 6 Tamil categories is sig-
nificantly anti-iconic and none of the Japanese categories is. On the other
hand, one Tamil and one Japanese category are significantly iconic. Thus
for the expressives we have clear evidence, albeit from a relatively small
sampling of categories, that words within narrow semantic categories ex-
hibit relative iconicity and for concrete nouns we have some evidence,
albeit from an even smaller and less rigorously selected sample, that words
within narrow semantic categories tend not to exhibit relative iconicity.
There are two rather striking exceptions to this latter conclusion among
the sample, one for Tamil and one for Japanese. For Japanese, the relatively
high form similarity within the category of internal organ nouns is clearly
related to the presence of the syllable zoo at the end of three of the words.
All three of these are early borrowings from Chinese (as is about one-third
of the Japanese lexicon), so this syllable probably behaves like a morpheme
for Japanese speakers (although the syllable that remains if zoo is deleted
probably does not). Thus, this set of words may not have satisfied our crite-
rion of mono-morphemicity. On the other hand, we have no explanation for
the high form similarity within the category of Tamil musical instrument
words.

9 Discussion and Conclusion

The main contribution of our study is the demonstration that it is possible to
quantify and measure relative iconicity, given agreed-on measures of pho-
netic and semantic similarity. Our more specific conclusions must be seen
as somewhat tentative because of the small sample of words from each lexi-
cal category and questions about how we performed meaning comparison.
These conclusions include the following:

4 Asterisks indicate significance at the .05 (*), .01 (**), .001 (***) and .0001 (****) levels.
If a Bonferroni correction is made, the significance cutoff rises: bars marked with a single
asterisk will no longer be significant.
 17 / ICONICITY IN EXPRESSIVES: AN EMPIRICAL INVESTIGATION

1. Expressives in Japanese and Tamil exhibit more relative iconicity

than concrete nouns.
2. As the size of an expressive semantic category gets smaller, the
likelihood of relative iconicity increases. If this effect proves to
hold for other languages, the communicative advantages that we
propose for iconicity may only apply for very narrow categories.
3. In very narrow semantic categories within concrete nouns, anti-
iconicity is also possible. If this result generalizes to other lan-
guages, the communicative advantages that we propose for arbi-
trariness or anti-iconicity may only apply for narrow categories.
In future work we will need to address both form and meaning com-
parison. Our form comparison algorithm may be missing aspects of per-
ceived similarity because it fails to take into account reduplication and we
also have only guessed at the appropriate settings for the small number of
parameters that govern the similarity measure, for example, at the relative
weight to assign to the onset and the nucleus within a mora and at the rela-
tive weight to assign to phonetic-feature differences within segments. What
is needed is experimental validation of the measure we are using.
Meaning comparison is clearly much more problematic. In order to se-
lect semantic categories and to select the words in the categories, we relied
on dictionaries and thesauri, but these sources rely in turn on the intuitions
of their authors. We are currently exploring ways of using subject similarity
ratings as an alternative.
We also will have to address several larger theoretical issues. One of
these concerns the whole notion of “meaning similarity”. On any account,
meaning is multi-dimensional and a person’s judgment of the similarity
between two words may be based on different dimensions or combinations
of dimensions, for example, whether one compares animals by their overall
SHAPE or their TAMENESS. Crucially, which dimensions come into play may
depend on the communicative context. Our preliminary investigations sug-
gest that certain dimensions may favor iconicity whereas others favor arbi-
trariness/anti-iconicity. For nouns, a dimension such as POLITENESS or
REGISTER may be more constraining for words that are maximally dissimilar
on that dimension but similar on all others (for example, feces and poop):
such words should be dissimilar in form. In future work, we will need to
examine how meaning similarity depends on context and whether certain
specific communicative contexts are more likely to lead to the effects we
predict for expressives and concrete nouns.
Lastly, one question we have barely touched upon is how composition
interacts with arbitrariness/anti-iconicity, e.g., in sets such as blackberry,
strawberry, blueberry. Here we find form similarity (a common morpheme)
for semantically related words, a relationship that runs counter to what we
would predict for concrete nouns (and leads to our anomalous finding for
Japanese nouns in the INTERNAL ORGAN category). It may be that in such
cases, addressees learn to use the morpheme overlap to help them recognize
the gross meaning category and to use the dissimilarity of the non-
overlapping portions to avoid confusion among subcategories within this
gross category. We plan to investigate this possibility using native-speaker
judgments of form similarity for mono-morphemic and poly-morphemic
pairs of words.

References
Chang, A. C. (1990). Gitaigo-giseigo bunrui youhou jiten. A thesaurus of Japanese
mimesis and onomatopoeia: Usage by categories. Tokyo: Daishuukan Shoten.
Childs, G. T. (1994). African ideophones. In Hinton, Nichols, & Ohala (Eds.),
Sound symbolism, pp. 179-204. Cambridge: Cambridge University Press.
Fenson, L., Dale, P.S., Reznick, J.S., Thal, D., Bates, E., Hartung, J.P., Pethick, S.,
& Reilly, J.S. (1993). The MacArthur Communicative Development Inventories:
User’s guide and technical manual. San Diego: Singular Publishing Group.
Fromkin, V. (1971). The non-anomalous nature of anomalous utterances. Language,
47, 27-52.
Gasser, M. (2004). The origins of arbitrariness in language. Proceedings of the An-
nual Conference of the Cognitive Science Society, 26.
Goldstone, R.L., Medin, D.L., & Halberstadt, J. (1997). Similarity in context. Mem-
ory & Cognition, 25 (2), 237-255.
Hinton, L., Nichols, J., & Ohala, J. (Eds.) (1994). Sound symbolism. Cambridge:
Cambridge University Press.
Luce, P.A. & Pisoni, D.B. (1998). Recognizing spoken words: The neighborhood
activation model. Ear & Hearing, 19, 1-36.
Ogura, T. & Watamaki, T. (1998). Research and development of early language
development inventory. Research project, Grant-in-aid for Scientific Research,
Kobe University (in Japanese).
Peirce, C. S. (1998). The essential Peirce: Selected philosophical writings: Volume
2. Bloomington, IN: Indiana University Press.
Sataasivam, M. (1966). Olikkurippakaraati (Dictionary of onomatopoeic expres-
sions). Madras: Paari Nilayam.
Saussure, F. de (1966). Course in general linguistics (tr. W. Baskin). New York:
McGraw-Hill.
Sethuraman, N. (ms.) Early language inventory: Tamil. Manuscript.
Sloutsky, V.M., & Fisher, A.V. (2004). Induction and categorization in young chil-
dren: a similarity-based model. Journal of Experimental Psychology: General,
133, 166-188.
Taub, S. F. (2001). Language from the body: Iconicity and metaphor in American
Sign Language. Cambridge: Cambridge University Press.